Anaphora and Semantic Interpretation: A Reinterpretation of

The Interpretive Tract. MIT Working Papers in Linguistics 25, 205-246U. Sauerland and O. Percus (eds.) © 1998 Irene Heim

Anaphora and Semantic Interpretation:A Reinterpretation of Reinhart's Approach*

Irene HeimMIT

Syntactic structure constrains the possible "coreference" relations between thenoun phrases in a sentence. Since the early days of generative syntax, we havelearned a great deal about the nature of these structural constraints. Much of thissyntactic theorizing has been possible with only a vague and informalunderstanding of what it really is in semantic terms that is being constrainedhere. Everyone agrees that it is not coreference in a literal sense, i.e., samenessof referents. But to specify in positive and semantically precise terms what it isinstead has turned out to be non-trivial. Those authors that have given seriousattention to the semantic import of syntactic constraints on so-called"coreference" have arrived at rather different conclusions.1

A particularly elegant and well-elaborated proposal in this regard is dueto Tanya Reinhart (1983a; 1983b). I ts central thesis is that only one type of"coreference" relation is syntactically represented and directly constrained byprinciples of grammar, and this is the well-understood relation of variablebinding in the sense of formal logic. Other semantic relations, in particular suchas may obtain between two referring NPs, are not even represented on anysyntactic level and can therefore not be directly licensed or ruled out bystructural conditions. They are regulated instead by an extragrammaticalprinciple which says, very roughly, that coreference is unavailable whenever thesame meaning can be conveyed by means of variable binding. We will get tothe details shortly.

The purpose of this article is two-fold: First, I will develop aformulation of Reinhart's proposal that is in certain respects somewhat moreexplicit than her own. Second, I will argue for a revision which, though it does *This paper was written in 1992 and distributed in the working paper series of theSeminar für Sprachwissenschaft of the University of Tübingen (SfS-Report-07-93). It isreprinted here in its original form (except for typographical corrections and updatedpublication dates in the bibliography).

Aside from its obvious debt to the work of Tanya Reinhart, the paper drawsheavily on insights that I obtained from the writings of James Higginbotham. AngelikaKratzer's explication of Reinhart's theory in her 1990 lecture notes also influenced mythoughts. James Higginbotham, Hans Kamp, Tony Kroch, Barbara Partee, Carl Pollard,Arnim von Stechow, and Hiroaki Tada offered important comments on earlierpresentations of this material, and Tanya Reinhart has been extremely generous with oralas well as written feedback for several years.1The following works are helpful in clarifying the problems and contain a representativesample of proposed solutions: Evans 1980; Higginbotham 1980 (note 1); Heim 1988(ch. III, 2.3); Soames 1990; Fiengo & May 1994 (ch. 1).

Irene Heim

206

leave practically all of Reinhart's substantive insights intact, leads to a theorywhich no longer fits the summary I just gave. In particular, the revised theoryimplies that bound variable anaphora does not enjoy a special status, butcoreference anaphora is syntactically represented in exactly the same way.

The reader is warned, first, that it is not the purpose of this article tosurvey or evaluate the competitors to Reinhart's approach which are alreadyfound in the literature. Aside from some scattered allusions, they will bedisregarded. Of course, if there already is a successful alternative to Reinhart'sapproach on the market, then this makes the present enterprise more or lessirrelevant. I do not believe that there is, but it would take a separate paper (orseveral) to explain just why not. Second, I will also disregard the numerouscriticisms that other authors have already put forward against Reinhart.2 Someof them, I believe, have been successfully countered or happen not to applyagainst my particular formulation of her ideas. Others remain unrefuted, andmost of those will threaten the revised theory I endorse no less than Reinhart'soriginal version. For instance, I inherit what are likely to be the wrongdescriptive generalizations about Weak Crossover3 and about ellipsis 4. If I amlucky, appropriate remedies for these and other shortcomings will notundermine my main points, but for all I know they might. Apart from theseobjections, which I am simply not competent to deal with, I neglect others formere reasons of space. In particular, I omit all discussion of Binding ConditionC, even though I defend claims that are not consistent with Reinhart's positionon this matter.5

1. Reinhart's theory

Reinhart's theory of the syntax and semantics of anaphoric relations is bestknown from her book Anaphora and Semantic Interpretation (Reinhart 1983a).My exposition mostly follows the more recent summary in Grodzinsky &Reinhart 1993 (henceforth G&R). I take the liberty of making some smalltechnical changes to suit my personal tastes and habits, but these should notdistort any of the substantive ideas.

Let us begin with the central assumptions about the derivation andwell-formedness of S-structure (SS) . We have free, optional indexing . Any NPmay, but need not, be assigned an index (a numerical subscript), and different

2See especially Lasnik 1989 (ch. 9) and references cited there.3For counterexamples and alternative proposals, see especially Higginbotham 1980 andStowell 1987.4Reinhart and the present work basically follow Sag 1980, which is problematic in lightof a number of more recent studies (see especially Dalrymple, Schieber, Perreira 1991,Kitagawa 1991, and Fiengo & May 1994.)5I am persuaded that Condition C is required in the syntax in order to predict thedistribution of bound variable construals for epithets, as shown, e.g., by Haïk 1984(204f.), Lasnik 1989 (ch. 9), and Higginbotham 1994. This point is independent of theargument that I have with Reinhart in this article. Once Condition C is reintroducedalongside A and B, most of what I say about Condition B below will probably carry overmutatis mutandis to C.

Anaphora and Semantic Interpretation

207

NPs in a sentence may receive the same or different indices. For instance, wecan assign (1) any of the following indexings (among others6), though (as wewill see below) some of these lead to derivations that are later filtered out as ill-formed or uninterpretable.

(1) Every boy called his mother.

a. every boy1 called his1 mother

b. every boy1 called his2 mother

c. every boy called his1 mother

d. every boy1 called his mother

e. every boy called his mother

One important filter consists of the Binding Theory conditions A and B, whichapply at SS:

(2) Binding Conditions:

A. An anaphor is A-bound in its GC.B. A pronominal is not A-bound in its GC.

(2) presupposes a lexical categorization into anaphors (in particular reflexivepronouns) and pronominals (non-reflexive personal pronouns). It alsopresupposes suitable characterizations of "Governing Category" (GC), for whichthe reader is referred to the syntactic literature. Moreover, it relies on thefollowing definitions, which in turn appeal to standard definitions of "A-position" and "c-command" (most of the time, it will not matter which particularversion).

(3) a. α binds β iff α c-commands and is coindexed with β.

b. α A-binds β iff α binds β and α is in an A-position.

None of (1a-e) happen to be filtered out by the Binding Conditions, but thereader is surely familiar with examples that would be.

From SS, a transformational derivation leads to Logical Form (LF) .The main operation of interest in this derivation is so-called Quantifier Raising(QR) . Contrary to what its name suggests, it applies optionally and freely to alltypes of NPs. (But again, derivations in which QR has failed to apply will oftenbe ruled out by yielding uninterpretable outputs.) Specifically, QR is assumedto apply in the following fashion: It replaces an indexed NP αi by a coindexedtrace, adjoins α (without the index!) to a dominating node, and prefixes thesister constituent of α with a lambda operator indexed i. Schematically: 6For instance, I neglected the additional option of indexing the NP his mother.

Irene Heim

208

(4) QR:[S ... αi ... ] => [S α λi[S ... ti ... ] ]

For example, the result of applying QR to every boy1 in (1a) is (1f).

(1) f. every boy λ1[t1 called his1 mother]

Our formulation implies that an NP needs to have an index in order to be able toundergo QR; this means, e.g., that every boy cannot be QRed in (1c) or (1e).This in itself presumably doesn't jeopardize the derivations, because quantifiersin subject position are straightforwardly interpretable in situ, there being nosemantic type mismatch.

LFs are then submitted to the following definition of "variable" and theassociated filter.

(5) a. An index is a variable only if it is(i) on a λ, or(ii) on a trace and bound by a λ, or(iii) on a pronominal or anaphor and A-bound.

b. All indices must qualify as variables.

(5) cuts down on the number of possible derivations quite considerably. For onething, it implies that all (overt) NPs apart from anaphors and pronominals, inparticular all quantifiers and proper names, must wind up without an index.This means they must either start out unindexed at SS, or else undergo QR andthereby transfer their index to the λ. (Which makes QR effectively obligatoryfor every boy1 in (1a,b,d).) From a semantic point of view, it makes sense not toallow indices on quantifiers and names: on standard assumptions, the meaningsof such NPs are completely determined by the lexical entries for the words inthem and compositional rules; an index has no conceivable semanticcontribution to make and would thus have to be ignored anyway if it werepresent.

(5) also implies that LFs cannot contain any free variables ("free" in thesense of standard logic book definitions). For instance, the result of QRing onlyevery boy1 in (1b) is not a legitimate LF, because the index 2 is not sanctionedby any clause of (5). (The only chance of rescuing this derivation would be byQRing his2 as well.)

(1) g. * every boy λ1[t1 called his2 mother]

(5) moreover incorporates a version of the Weak Crossover prohibition, bydisallowing locally A- bar -bound pronouns. So the derivation (1a)/(1f) is well-formed, because the pronoun his1 winds up A-bound by the QR-trace t1, but itscounterpart with subject and object reversed would not be:

(6) a. SS: his1 mother called every boy1

b. LF: * every boy λ1[his1 mother called t1]


209

The index on his in (6b) fails to count as a variable, because its only binder isthe λ rather than an A-position.

Finally, LFs are semantically interpreted . I can be informal about thishere, because how it works is mostly obvious. λ's, of course, are functionalabstractors. Variables in the sense defined in (5) are treated like variables inlogic (with occurrences of the same numerical index counting as occurrences ofthe same variable, and occurrences of distinct numerical indices as occurrencesof distinct variables). A constituent that bears an index simply inherits theinterpretation of its index. For this reason, it is harmless and natural to refer tothe whole indexed NP as a "variable", though strictly speaking, the variable isjust the index.

Pronouns without indices are deictic and the utterance context has toprovide referents for them. Apart from sortal restrictions due to the pronoun'sgender, number, and person features, this reference assignment is a result ofinteracting pragmatic factors, including salience and overall plausibility.Moreover, it is constrained by the following principle, which constitutes themost distinctive ingredient of Reinhart's approach.

(7) Coreference Rule:

α cannot corefer with β if an indistinguishable interpretation can be generated by (indexing and moving β and)7 replacing α with a variableA-bound by the trace of β.

The next section is entirely devoted to illustrations of (7), which will also serveto clarify some of the concepts it employs, notably "interpretation" and"indistinguishability".

2. Reinhart's Coreference Rule applied to examples

This section parallels the discussion of the Coreference Rule in G&R, section2.3. In particular, my example groups (ii) - (v) are all taken from their list, withone systematic alteration: Since I am not dealing with Condition C effects at allin this paper, I have replaced all apparent Condition C violations by similarexamples that look as though they violate Condition B.

2.1. Group (i): basic casesLet's look at three primitive examples containing the proper name John and amasculine singular pronoun.

(8) John saw him .

(9) John saw his mother.

(10) His mother saw John . 7The parenthesized part of the instructions can be skipped if β was already QRed in theoriginal structure.

Irene Heim

210

In each case, we are interested in the possible relations between theinterpretations of the name and the pronoun. This breaks down into twoquestions: First, can the two stand in a variable binding relation? Second, canthey corefer? The predictions turn out to be the following: (8) allows neitherbinding nor coreference; (9) allows binding but not coreference; and (10) allowscoreference but not binding. Here is how they are arrived at:

When we ask whether the him in (8) could be a variable bound by John,we can't, of course, mean this quite literally; proper names are not variablebinders, after all. What we really mean is whether the pronoun could be avariable bound by the λ that arose from QRing the name. So the question ofwhether binding is possible in (8) turns on the well-formedness of the followingderivation.

(8) a. SS: * John1 saw him1

b. LF: John λ1[t1 saw him1]

Though all indices in the LF (8b) qualify as variables and there is no obstacle tointerpretability, this derivation is already filtered out at SS by Binding ConditionB.

Could the two NPs corefer? For this we would need an LF as in (8c)(trivially derived from an identical SS in which no NP was indexed) and anutterance context that furnishes the reference assignments indicated by thepointers underneath.

(8) c. LF:John saw him ↓ ↓

j j

I will make use of this notation to specify utterance contexts throughout thepaper: Each referring NP in the LF is connected by an arrow to its contextuallysupplied referent. The lower-case letters stand for individuals out there in theworld, with each letter representing a unique individual and each individualrepresented by a unique letter.

There would be nothing wrong with the interpretation indicated in (8c),if it weren't for the Coreference Rule. This rule instructs us to look for analternative LF that results by certain specified alterations from that in (8c) and tomake sure that it wouldn't yield an indistinguishable interpretation. A potentialsuch alternative happens to be the LF we already saw in (8b), set in this context:

(8) d. LF:John λ1[t1 saw him1] ↓ j

This differs minimally from (8c) in just the way that (7) instructs us: Of our twoNPs in (8c) that were candidates for a coreferring pair, the first (John) has beenindexed and QRed and the second (his) has been replaced by a variable (his1)that is A-bound by the trace of the former. It is an interpretable LF,notwithstanding the fact that it (because it would have to derive from a


211

Condition B violation at SS) is not part of a well-formed derivation for anyEnglish sentence.8

Now we must determine whether there is any difference in theinterpretations of (8c) and (8d). Suppose for now this is to be done bycomparing the propositions expressed by each LF in its indicated context . Itturns out that these are the same: (8c) is true in any world where j saw j, and sois (8d). So (8d) represents an indistinguishable interpretation from (8c), andtherefore the Coreference Rule says that (8c) is not a permissible interpretationfor (8). In short, the option of coreference in (8) is preempted by the existenceof (8d).

Let's turn to (9).


Here, binding is evidently predicted possible, the following derivation beingwell-formed on both levels.

(9) a. SS: John1 saw his1 mother

b. LF: John λ1[t1 saw his1 mother]

But coreference is not. A coreferential interpretation would look like this:

(9) c. LF: John saw his mother ↓ ↓ j j

But under the Coreference Rule, this is preempted by (9d) (= (9b) plus acontext).

(9) d. LF: John λ1[t1 saw his1 mother] ↓ j

Both of these express the proposition that j saw j's mother.(As G&R acknowledge in footnote 13, it might be preferable to predict

this example to be ambiguous between a bound and a coreferential reading.They suggest that this could be accomplished by a revision of (7) that confinesits application to those examples which involve prima facie violations ofCondition B, in a sense they make precise. I will adopt a similar proposalbelow, but set the issue aside for the time being.) 8I have chosen to read (7) in such a way that the potential alternative structures to beconsidered in applying this principle need not be part of complete grammaticalderivations. Alternatively, one might impose this further requirement, in which case thehim in (8c) would have to be replaced by a himself1 in (8d). Most renditions of Reinhart'sproposal seem to assume the latter. It doesn't seem to make any difference for the casesconsidered here, but see footnote 12 of G&R (and apparently Reinhart (1991a), which Ihaven't seen).

Irene Heim

212

In (10), we have the reverse prediction.

(10) His mother saw John .

Binding is out, because (as we already saw with (6b)), the requisite structurecontains an index on his1 that fails to qualify as a variable.

(10) a. * John λ1[his1 mother saw t1]

But coreference is, for this very reason, allowed. (10b) depicts the relevantinterpretation.

(10) b. LF:his mother saw John ↓ ↓ j j

This is not preempted by any other structure. Why not? Because the closest wecan come to constructing a potential competitor according to the specificationsof (7) is to index and QR John and coindex his with its trace, but then we haveprecisely (10a), where his1 is not a variable.

These three examples should have clarified some mechanical aspects ofthe Coreference Rule. They also gave us the opportunity for a first stab atelucidating the notion of indistinguishable interpretations, but we will soon seethat there is more to this notion than we have so far uncovered.

2.2. Group (ii): examples with only(11) illustrates another type of example which has the superficial appearance ofa Condition B violation and which Reinhart cites in support of the CoreferenceRule. (I won't talk about analogous cases with other focussing particles such aseven.)

(11) (Everybody hates Lucifer.) Only he himself pities him .

Why is (11a), with coreference, an available interpretation and not preempted bythe binding-alternative (11b)?

(11) a. only he himself λ1[t1 pities him] ↓ ↓ l l

b. only he himself λ1[t1 pities him1] ↓ l


213

Well, that's obvious. We don't need to cast around for subtle differences here.These two LFs have manifestly different truth conditions. (11a) says thatnobody besides Lucifer pities Lucifer , whereas (11b) claims that there are noother self-pitiers .

Actually, a closer look at (11b) reveals a potential problem with thisapplication of (7). Notice that what corresponds to the α and β of (7) here arethe NPs him and he himself respectively. It is these two, after all, whosecoreference is to be licensed. When we construct the potential competingstructure according to the specifications of (7), we are supposed to coindex αwith the trace of β . But the trace t1 with which we coindexed him1 in (11b) isnot really the trace of he himself , but rather the trace of only he himself . That isa different NP, and not a referring one, hence not a possible choice for β in thefirst place.

In short, if this is indeed the way in which G & R intend theCoreference Rule to apply to this example, then they must somehow beassuming that for the purposes of (7), t1 in (11b) counts as a trace of he himself.That seems a little bit hokey, but there is something to be said for it. I willreturn to the matter in section 5.3.3 below.

2.3. Group (iii): identity under debateCoreference is systematically possible in Condition B configurations when weare dealing either with explicit identity statements or with other utterances indiscourse contexts where the identity of the referents is unknown or at issue. Arepresentative example is the second to last sentence in (12).

(12) A: Is this speaker Zelda?

B: How can you doubt it? She praises her to the sky. No competingcandidate would do that.

Nothing is wrong with this if the woman in question indeed is Zelda and thepronouns thus corefer. Nothing is even wrong if the speaker knew this all alongand makes no secret of it. Reinhart's Coreference Rule is meant to throw lighton this well-known phenomenon.

Needless to say, variable binding is ruled out in the familiar way byCondition B at SS. Coreference amounts to the following interpretation:

(12) a. she praises her to the sky ↓ ↓ z z

We must show that (12a) is not preempted under the Coreference Rule by anyother structure. A potential competitor with the right linguistic shape would be(12b).

(12) b. she λ1[t1 praises her1 to the sky] ↓ z

Irene Heim

214

If this doesn't qualify to preempt (12a), it can only be because its interpretationis distinct. Is it?

Not if we just compare the propositions expressed. They are one andthe same, viz. that z praises z to the sky. So if we continue to interpret theCoreference Rule in the way we did in the previous sections, it founders on thisexample. Alternatively, we must look for a more suitable notion ofindistinguishable interpretation that will make it work right. The latter is, ofcourse, what Reinhart intends.

It is commonplace in the philosophical literature (especially on identitystatements) to distinguish the proposition expressed by an utterance from itscognitive value .9 When the utterance contains referring terms 10, the propositionexpressed depends only on their referents, but the cognitive value depends alsoon the way these referents are presented. In the context of our example (12), forinstance, the person z presents herself to the interlocutors in two differentguises: First, they have a current visual impression of her, standing on theplatform over there and speaking. Second, they carry in their memory an entrywith various pieces of information about her, including that she is called"Zelda".

Now when it comes to processing our sentence she praises her in (12),what intuitively goes on seems to be this: Each of the two pronouns connects toits referent z via one of these two guises. She, because of a perceived anaphoriclink to the subject this speaker of the preceding sentence, associates with thevisual impression; and her, through its link to the postcopular NP Zelda,activates the memory entry. Therefore, the cognitive value of the sentence shepraises her to the sky for the hearer in this context is the proposition thatwhoever causes the visual impression in question praises whomever thepertinent memory entry represents. This is rather a different proposition fromthe one the sentence expresses, viz. that z praises z.

More importantly for our present purposes, it is also different from thecognitive value of the potential competitor in (12b). How so? Because (12b)contains only one referring NP (she) and this will pick out its referent z via oneof the two guises salient in this context. Presumably this is the same one as forthe she in (12a), namely the visual impression. The cognitive value of the LF in(12b) thereby comes to be the proposition that whoever causes this visualimpression praises herself—clearly a different proposition (with differenttruthconditions) from the cognitive value of (12a) as described above.

The level of cognitive values thus seems to be more appropriate thanthat of propositions expressed when it comes to distinguishing interpretations inthe sense of Reinhart's Coreference Rule. In this respect, our little diagramsrepresenting interpretations have been misleading or at least incomplete. Anutterance context for referring pronouns doesn't just supply these pronouns withreferents. Rather, it supplies them with guises , and these in turn happen to beguises of something, namely the referents. The diagrams should thus contain an 9While some distinction along these lines is commonplace, many details and, of course,the terminology vary from author to author. (The term "cognitive value", for instance,comes form the introductory passage of Frege 1892, as translated by Max Black in Black& Geach 1952.) For a thorough introduction and overview, see Haas-Spohn 1995.10More accurately: directly referential terms in the sense of Kaplan 1989.


215

additional intermediate layer for the guises. (I use capital letters 'F', 'G', ... forguises, with each letter standing for a unique guise and vice versa.) (12c)replaces (12a), and (12d) replaces (12b).

(12) c. she praises her to the sky ↓ ↓ F G ↓ ↓ z z

d. she λ1[t1 praises her1 to the sky] ↓ F ↓ z

Semantically speaking, a guise is an individual concept, i.e. a function fromworlds to individuals. For instance, the visual impression alluded to above(represented by 'F' in (12c,d)) can be viewed as assigning to each possible worldw the individual that it is an impression of in w. In the actual world of theutterance, this is z, but in other worlds it may be someone else. The cognitivevalue of (12c) is the proposition that is true in any w where F(w) praises G(w) tothe sky; that of (12d) is the proposition that is true in any w where F(w) praisesF(w) to the sky.

I will not formalize this any further, and in subsequent sections, I willeven fall back into the simplifying pretense that utterance contexts matchpronouns simply with referents. But before it's safe to do so, we should makesure that the stories we told about (8) - (10) in the last section haven't collapsedin the light of our refined notion of indistinguishable interpretation.

For instance, we took coreference between John and him to be ruled outin (8) because (8d) preempted (8c). It did so, we said, because it expressed thesame proposition. But now we have seen that sameness of propositionexpressed is not a sufficient condition for indistinguishable interpretations. Soour argument re (8) is no longer conclusive. We should have established that(8d) has the same cognitive value as (8c). Can this stronger argument be made?Well, it can, if we bring out and exploit a tacit assumption about the example,namely that it was meant to be judged either out of context, or in some sort ofrun-of-the-mill context, say a conversation about John in his absence. In thatkind of ordinary setting, there wouldn't be multiple salient guises of j, but justone (presumably the memory entry under his name), and the context wouldassign that one to both the name and the pronoun in (8c). And in that case, thecognitive values of (8c) and (8d) coincide, as desired. (Even if j happened to bepresented in two ways—say, he was visible in the distance during theconversation about him—, this wouldn't suffice for the pronoun to automaticallylink to him via a different guise than the name. For that to happen, the contextmoreover has to contain appropriate clues that this is the intendeddisambiguation.)

In short, we must qualify our earlier conclusion about (8): It doesn'treally follow from the grammar and the Coreference Rule alone that (8c) is

Irene Heim

216

unavailable; it only follows under additional assumptions about the context,which imply that there is only one salient guise of j. But come to think of it, thisqualification is a good thing. It is a fact, after all, that even (8) could be usedwith John and him coreferring if it were placed in an appropriately contrivedcontext.

There is much more to be said about the pragmatic conditions underwhich contexts make two distinct guises of the same referent readily enoughavailable11, but let's move on.

2.4. Group (iv): when structured meaning mattersEvans (1980) emphasized a type of example that G&R likewise bring up inillustration of the Coreference Rule. It will point us to yet another aspect of thenotion of indistinguishable interpretations. (The examples are again not exactlyEvans's or G&R's, but adaptations thereof to Condition B configurations.)Consider the last clause of (13).

(13) (You know what Mary, Sue and John have in common? Mary admiresJohn, Sue admires him, and) John admires him too.

Despite the Condition B environment, one sort of gets away with him referringto John here. Apparently there is something about this particular precedingdiscourse that makes it possible. What exactly is it and how does theCoreference Rule predict it to matter?

The coreferential interpretation in question is (13a), and for somereason it is not preempted by (13b).

(13) a. ... and John admires him ↓ ↓ j j

b. ... and John λ1[t1 admires him1] ↓

j

11In Heim 1988:315 - 320, I proposed one concrete restriction: No context ever assignsdistinct but presupposedly coreferential guises to any pair of NP-occurrences.(Definition: Guises F and G are presupposed to corefer in context c iff F(w) = G(w) forevery world w that conforms to the shared presuppositions of the discourse participants inc.) This means that reference to the same object via two distinct guises is possible onlyas long as the speaker still treats it as an open question whether indeed the same object isbehind these two guises. Once this is taken for granted (more accurately: presupposed inthe sense of Stalnaker 1979), only one guise is available. (This may be the result of"collapsing" two previously available guises; in technical terms, the result of collapsing Fand G is F restricted to the set of worlds on which it coincides with G.) I still think thisproposal is defensible, though there are non-trivial issues to sort out (see, e.g. Landman1986:104 - 105 for critical discussion). For the purposes of the present article, however, Ineed not commit myself.


217

Why not? The propositions expressed are again identical, and there also don'tseem to be two different guises of j in this context that would lead to differentcognitive values. What then is the difference? It is, I think, the fact that twodifferent properties are predicated of j in (13a) and (13b): the property P ofadmiring j in (13a), and the property Q of admiring oneself in (13b). In otherwords, (13a) and (13b) express different structured propositions 12, <j,P> and<j,Q>, even though they express the same unstructured proposition P(j) = Q(j) =that j admires j.

Fair enough, but why does this suddenly matter? We could haveobserved the same thing about all our earlier examples: there, too, the structured propositions differed, even where we concluded that the interpretations wereindistinguishable. Evidently, differences in structured meaning do not alwaysmatter. There has to be a special reason when they do. The special reason inthis case emerges when we analyze the preceding sentences: The speaker startsout with a promise to tell what Mary, Sue, and John have in common. So weexpect him to mention a property that each of the three has. One particularlydirect and rhethorically effective way of doing this is to utter three predicationsin which the same property is predicated of each of the three people in a row.Now the first two predications, concerning Mary and Sue, were as follows:

(13) c. Mary admires John, Sue admires him, ... ↓ ↓ ↓ ↓

m j s j

Each of these predicates of its subject the property P of admiring j. Now if wecontinue by (13a), we get a third predication of that same property. (13b), onthe other hand, would break the pattern and predicate a different property of jthan of the previous two people. It would still, of course, give us indirectinformation about what the three have in common: we could determine theshared property by a simple bit of deduction. But (13b) is more explicit, andthis, I submit, makes it beat out its competitor here and avoid being preempted.A similar story applies to (14). This utterance, we are to imagine spoken by alogic tutor.

(14) Look, if everyone hates Oscar, then it surely follows that Oscar (himself) hates him .

The intended interpretation is (14a).13

(14) a. if everyone hates Oscar, ... Oscar hates him ↓ ↓ ↓

o o o

Why isn't it preempted by (14b)? 12Structured propositions (and other types of structured meanings) have been put tovarious uses in semantics; see Cresswell & Stechow 1982 for a recent example.13Since subject quantifiers are interpretable in situ, I didn't bother to index and QReveryone, but of course it wouldn't have hurt to do so.

Irene Heim

218

(14) b. if everyone hates Oscar, ... Oscar λ1[t1 hates him1]

↓ ↓ o o

Well, what the logic tutor is apparently trying to get across to the student is howto apply the law of Universal Instantiation. (14a), as it happens, is a pureillustration of that law: the predicates following everyone in the antecedent andOscar in the consequent denote the same property, that of hating Oscar. (14b),by contrast, has the property of hating Oscar in the premise, but a different one,that of hating oneself, in the conclusion. Of course, it is likewise a validinference. But its validity relies on more than just Universal Instantiation; itcollapses two inference steps (U.I. and λ-conversion). And that isn't optimaldidactic practice in this context.

So once again, a difference in structured meaning alone matters enoughto allow the interpretations (14a) and (14b) to count as distinct. But here as inthe previous example, this is due to very special circumstances: Logic teachershave a professional duty to care not just about what proposition a sentenceconveys, but about how that proposition is built up from parts. For mostordinary conversational purposes, however, the net message is all that counts,and so we were right to disregard mere differences in structured meaning withour earlier examples, and to disregard them again with most of the ones below.

2.5. Group (v): Lakoff's example14

Finally, why does one get away with utterances like (15), due to Lakoff (1972:639)?

(15) I dreamt that I was Brigitte Bardot and I kissed me .

The two underlined pronouns, being both first person, supposedly can't help butcorefer:

(15) a. ... and I kissed me ↓ ↓

g g

('g' for George Lakoff, the speaker of (15).) But this interpretation ought to bepreempted by (15b).

(15) b. ... and I λ1[t1 kissed me1] ↓ g

14This section is substantially changed from the previous version of this paper, partly inresponse to questions raised by Higginbotham (p.c.) and Reinhart 1991b. Thanks toArnim von Stechow for reminding me of Stechow 1982. (McKay 1991 also looksrelevant, but I only received it when I was almost finished.)


219

Unless, again, we can argue that the latter has a distinct interpretation. That thisshould be the case is made plausible by the observation that a minimallydifferent sentence, whose grammar forces the variable binding we see in (15b),would actually describe a different dream:

(16) I dreamt that I was Brigitte Bardot and I kissed myself .

(16) describes a dream about self-kissing, (15) doesn't.15

Suggestive though this is, it is not so easy to show concretely how theinterpretations of (15a) and (15b) differ. The propositions expressed under thereference assignments depicted in (15a,b) are certainly the same. The structuredmeanings are different, but it is not evident why this should matter in thiscontext; at least there is no reason of the kind we found with (13) and (14)above. Are the cognitive values distinct, then? For all we know so far, thiscould only be if George Lakoff, as he utters (15), is somehow presented to hisaudience in two separate guises, and that doesn't seem to be the case here either.So where could the difference possibly lie?

To overcome this puzzle, we have to be a bit more precise on how suchexamples are semantically interpreted. A suitable analysis of pronouns inattitude reports, including an explicit application to (15), is found in Stechow1982. Adapting his analysis (with some inessential alterations), let's begin bypositing a more articulate LF representation for attitude complements than wetook for granted in (15a,b). Following Quine 1956, believe, dream , and otherattitude verbs are logically 3 -place predicates: the basic notion is for a subject tobelieve (dream) something of something . The third argument (the " res "-argument) does not correspond to a surface constituent, but it is present at LFand may be filled there by material moved out of the complement clause.16 Forinstance, a simplified version of (15) (omitting the complement's first conjunct)may have the following LFs, among others. (The corresponding SSs that theseshould derive from in Reinhart's framework—all well-formed by the BindingConditions—appear in parentheses underneath.)

15Actually, the data are not quite so simple. The choice of the reflexive seems to becompatible with both readings; at least this is my intuition about similar Germansentences:(i) Hans soll sich mal vorstellen, der Lehrer zu sein und sich/?ihn als Schüler zu

haben.'Hans should imagine being the teacher and having himself/him as a student.'

The variant with the pronominal ihn is unambiguous as predicted by the analysis I willsketch, though a little marginal (see below on what the marginality might be due to.) Thereflexive sich, however, also allows the pragmatically preferred reading, according towhich Hans imagines teaching Hans (rather than self-teaching). I have no account forthis reading. The analysis sketched in the text predicts only the self-teachinginterpretation here. Further research is required.16Unfortunately, the syntactic distribution of de re construals doesn't exhibit theproperties of movement. An "in situ" approach of the type that has proved successful forassociation with focus would therefore be more appealing. The techniques for this are inprinciple well worked out (see e.g. Rooth 1985) - except (to my knowledge) for theinteraction with variable-binding, which happens to be crucial in the present application.

Irene Heim

220

(15) c. I dreamt λ1[t1 kissed me] [ I ]

(SS: I dreamt I1 kissed me )

d. I dreamt λ1[I t1-ed me] [ kiss ]

(SS: I dreamt I kiss1ed me )

e. I dreamt λ1,2[t1 kissed t2] [ I, me ]

(SS: I dreamt I1 kissed me2 )

f. I dreamt [ I kissed me ] [ ]

(SS: I dreamt I kissed me )

The "res-movement" that creates these LFs is like QR insofar as it involves λ-abstraction, but in some other respects operates quite differently: First, it isn'tChomsky-adjunction, but substitution into a kind of argument position. Second,it may affect phrases other than NPs (as in (d)). Third, it can apply multiply (asin (e)). (If pure de dicto readings are to be covered as a special case, the res-slotmay also be left empty, as in (f).)

The semantic interpretation of the verb and its two internal arguments isas follows: The left argument, denotes (as the λ-notation implies) a property ofn-tuples (n≥0, depending on the number of indices on the λ). The res-argumentdenotes an n-tuple made up of the denotations of the phrases in it. Theinterpretation of the verb is relative to a special contextual parameter, an n-tupleof acquaintance relations in the sense of Lewis (1979). More precisely:

(17) An LF of the form [α β γ [δ1,...,δn] ], where β is an attitude verb and n≥0,requires an utterance context c which furnishes, for each i = 1, ..., n, anacquaintance relation Dci.

To complete the analysis, take dream to mean 'believe in one's sleep'. In a givencontext c, dream then denotes the following function fdream, c:

(18) fdream, c(P)(<a1,...,an>)(b) = 1 iff

(i) b uniquely bears Dc1, ..., Dcn to <a1,...,an>, and

(ii) b is asleep and self-ascribes the property of uniquely bearingDc1, ..., Dcn to some n-tuple of individuals satisfying P.

Now back to our example (15). Stechow proposes that (15) under the intendedreading has the LF-representation in (15c).17 The utterance context under 17It might actually be more accurate to assume (15e), with both pronouns res-moved.This would come to exactly the same reading if the second acquaintance relation suppliedby the context happened to be that relation which each individual bears to George Lakoff,


221

consideration (as before) assigns g as the referent to all three pronoun-occurrences, and (by (17)) moreover supplies an acquaintance relation D. Hereis another picture to summarize these aspects of LF and context. (The pointerunder the verb indicates the contextually supplied acquaintance relation.)

(15) g. I dreamt λ1 [t1 kissed me] [ I ]↓ ↓ ↓ ↓g D g g

What is D? It is simply identity; this, at any rate, seems to be the choice thatyields the intuitively salient meaning. What the utterance asserts, then, is that gself-ascribes in his sleep the property of kissing g. (The choice of identity for Damounts to a so-called de se reading for the embedded subject; as Stechownotes, this is usually the preferred reading for pronouns coreferential with thehigher subject.18)

Now, I think, we can finally see how G & R intend the CoreferenceRule to apply to this example. The me in (15g) is a referring pronoun thatcorefers with the other two 1st-person pronouns in the sentence, in particularwith the embedded I that originates from a c-commanding position in its GC.Why is this allowed?19 It would not be, (7) predicts, if (15g) meant exactly thesame thing as the following competitor with a bound variable:

(15) h. I dreamt λ1[t1 kissed me1] [ I ]↓ ↓ ↓g D g

In (15h), me has been replaced by a variable (me1) A-bound by t1, the trace of I,just as (7) demands. (The only difference compared to previous applications of(7) is that the trace in question is a res-movement-trace, rather than a QR-trace).(15h) has demonstrably different truthconditions from (15g): Given the samevalue for D (identity), (15h) claims that g in his sleep self-ascribed the property i.e. to that world-mate who is the actual Lakoff's counterpart by match-of-origins. It isdoubtful, however, that Lakoff knows enough about his origins to dream such a thing; seeLewis 1984 for discussion. An advantage of (15e) over Stechow's choice (15c) would bethat the second pronoun could then be interpreted w.r.t. a similar, but probably morerealistic, acquaintance relation, say, the relation that x bears to y iff y is a world-mate ofx's who has the history and permanent characteristics that the actual George Lakoffknows himself to have. But I will disregard this complication.18For more discussion of de se readings and their status as special cases of de rereadings, see Lewis 1979, Chierchia 1991, Higginbotham 1989, and Reinhart 1991b.19Stechow also asks himself why (15) is not a Condition B violation, but his answer isnot quite sufficient for our present purposes. According to him, (15g) is okay because theGC for me here is t1 kissed me, which does not contain a coreferential c-commanding NP(in fact, it contains no other referrring NP at all). But if this were good enough, whycouldn't we rescue every Condition B violation simply by QRing the offendingantecedent (e.g. as in John λ1[t1 saw him])? At any rate, it is not good enough forReinhart, whose theory we are assuming here. Under her assumptions, only looking at(15g) itself is not enough to license it; we must also consider potential competingstructures.

Irene Heim

222

of self-kissing (the meaning earlier observed in sentence (16)). Hence (15h)does not preempt (15g) and so (15g) is licensed.

3. Reference isn't special

Up to now I have merely tried to explain Reinhart's analysis. If I have gonebeyond plain repetition from G & R and her earlier publications, it was only toflesh out details in the semantic analysis of certain examples, but not to addanything that wasn't there yet, at least between the lines. In this section, I willbegin to disagree.

Let us look again at our fourth group of examples, the ones wherestructured meaning mattered ((13) and (14)). Reinhart's account of themcaptures an important intuition: These examples are licensed because of acontextually important aspect of their meaning that would get lost if they werereplaced by their bound-variable counterparts. As such they call for a rule of thekind of Reinhart's Coreference Rule, which essentially involves a comparisonbetween the meanings of two competing structures.

I still have a quibble, however: Is the phenomenon illustrated by (13)and (14) really peculiar to referring NPs? Is it only referential pronouns that weare sometimes allowed to use in unusual ways when a conversational purposejustifies it? Come to think of it, such a limitation wouldn't be particularlyplausible to expect in an essentially pragmatic principle of this sort. Andindeed, once we start looking for the relevant examples, it isn't well supportedempirically either.

Recall, for instance, our logic tutor and his excuse for the coreferentialuse of Oscar and him in (14). Once we let him get away with (14), are we reallygoing to put our foot down when he goes on as in (19)?

(19) ... And, of course, this doesn't just hold for Oscar, but for any arbitraryman: If everyone hates a man, then that man himself hates him .

My point, of course, is that the last sentence in (19) is a donkey-sentence and thetwo underlined NPs are donkey-anaphors, hence not referring terms. Under oneanalysis of donkey-sentences, they are plainly bound variables, i.e., the LFshould be something like (19b), with a silent adverb of quantification equivalentto a restricted universal quantifier (here abbreviated as "∀").20,21

20Under an alternative (E-Type) approach to donkey-sentences, the donkey anaphors arenot plain bound variables, but descriptions of some kind, and the silent universalquantification is not directly over men, but over cases. (See Neale 1990, Heim 1990, andothers for recent discussion.) Still, the donkey-anaphors are not referring terms underthat alternative either, because they would contain a bound case-variable, and so my mainpoint goes through all the same.21It is not entirely obvious at this point how this LF is to be derived and licensed. Onerevision to the present system that is surely required, once we bring in donkey-sentences,is that indefinites on the one hand and demonstratives (and other complex definites) onthe other be allowed to count as variables, along with pronouns and anaphors. Thedefinition of variables also needs extending for donkey anaphors that are ordinarypronouns, because these need not be A-bound. A suitable pair of added clauses might be:


223

(19) a. SS: * if everyone hates [a man]1,

[that man himself]1 hates him1

b. LF: ∀1 [ if [a man]1 [everyone hates t1] ]

[ [that man himself]1 hates him1 ]

But this cannot be generated under Reinhart's assumptions: the derivation wouldbe blocked by Condition B at SS. There is no provision for licensing ill-formedderivations like this, and the Coreference Rule is simply not pertinent here at all.I take it that this is not satisfactory: (19) should be predicted to have prettymuch the same status as (14), with the choice of him over himself allowed forthe very same reason. Admittedly, (19) is (even) more contrived than (14) andhighly questionable from a pedagogical point of view. But it seems to me thatits added complexity is sufficient to account for what degradation we perceivehere. There isn't a qualitative contrast that would justify the special status ofreferring pronouns that Reinhart's Coreference Rule in its present form grantsthem.

And while you are in the appropriately contrived mind-set, consider abound variable variant of our other "structured meaning" example, (13):

(20) Somebody said that what he had in common with his siblings was that hissister admired him, his brother admired him, and he (himself) admiredhim.

Again, it may not be the most natural English sentence, but the judgment isn'tsuch as to warrant a fundamental disparity between referential and boundvariable pronouns. My conclusion, therefore, is that Reinhart's CoreferenceRule captures the right intuition of what makes utterances like (13) and (14)possible, but it should be made a little more general so that it covers (19) and(20) as well.

Analogous points could be made about some of the other groups ofexamples, but are more easily obscured there by murky technical details.Recall, for instance, the only-example, whose intended interpretation is repeatedhere:

(11) a. only he himself λ1[t1 pities him] ↓ ↓ l l

This, too, has bound-variable counterparts with much the same intuitive status.For instance, (21) allows the reading in (21a).

(iv) or on an indefinite in the restriction of a QAdv and A-bar-bound by it,(v) or on a definite in the nuclear scope of a QAdv and A-bar-bound by it.

Irene Heim

224

(21) Every devil knows that only he himself pities him.

a. ∀x[devil(x) → know(x, that ∀y[y≠x → ~pity(y,x)])]

In Reinhart's system, this reading must have a derivation such as (21b,c).

(21) b. SS: every devil1 knows that

[only [he himself]1]2 pities him1

c. LF: every devil λ1[t1 knows that

[ only [he himself]1 λ2[t2 pities him1] ] ]

Perhaps this is okay because no well-formedness constraint at either level rulesit out (in particular, the SS doesn't violate Condition B, because the adjoinedonly blocks A-binding of him1 by [he himself]1). It is then not directlyproblematic for Reinhart's theory. But isn't it a little strange if the explanationfor the acceptability of reading (21a) in (21) is so completely unlike theexplanation that was given for the acceptability of reading (11a) in (11)? Thelatter involved comparison with potential preemptors under the CoreferenceRule, whereas the former relies entirely on considerations of syntactic well-formedness. We may suspect that a generalization is being missed here. Itwould require some work to turn this suspicion into a real objection, but since Iwon't let it carry the burden of my argument, I can afford to stop short of thishere.

The Lakoff-example likewise has bound-variable cousins, for instance(22).

(22) Not only I dreamt that I was Brigitte Bardot and I kissed me.

The relevant reading here is the one where (spoken by g) it says that there wassome x≠g such that x self-ascribed the property of kissing x (not to be confusedwith the property of self-kissing). This seems no less acceptable than thereading we have discussed for (15). But again, the explanation in Reinhart'sframework for why (22) allows such a reading cannot be anything like theaccount given above for (15). Presumably, (22) has this reading because itallows a derivation terminating in the following LF.

(22) a. not only I λ1[ t1 dreamt that ... λ2[t2 kissed me1] [ I1 ] ]

It is not so clear at this point how such an LF is derived: what exactly is the SSand why doesn't it violate Condition B? One way or another, these details mustbe sorted out if the reading in question is to be generated. I will offer a concretesuggestion below, but my present point doesn't depend on it. It is simply that,whatever the details of an account of (22) in Reinhart's framework may turn outto be, the Coreference Rule will not play any role in it. It couldn't, because thereis no coreference to be licensed (only a certain pattern of variable binding ). Soonce again, what to the naive observer appears to be just a more complicated


225

instance of the same phenomenon receives a qualitatively different explanation -- not, as it stands, a legitimate objection, but grounds for more suspicion.

The moral of this discussion seems to be that we should look for a moregeneral version of Reinhart's Coreference Rule, one that will cover boundvariable examples like (19) along with their simpler cousins involvingreferential NPs, like (14). What we would like to replace (7) with is somethinglike the following:

(23) Coreference-or-Cobinding Rule:

α cannot corefer or be cobound with β if an indistinguishableinterpretation can be generated by (indexing and moving β and) replacingα with a variable A-bound by the trace of β.

This formulation presupposes a distinction between α and β being cobound onthe one hand and α being bound by (=cobound with the QR-trace of) β on theother. (If "cobound" and "bound by" just meant the same thing, (23) would notmake sense, because then the cobound interpretation would be necessarilyindistinguishable from the one we are instructed to compare it to.) Thedistinction is clear enough. For instance, [that man himself]1 and him1 arecobound in our LF (19b) above, because each is bound by ∀1. If the latter wereto be bound by the former, the structure would have to look different, namelylike this:

(19) c. ∀1 [ if [a man]1 [everyone hates t1] ]

[ [that man himself]1 λ2[t2 hates him2] ]

The consequent clauses of (19c) and (19b) are logically equivalent, but theydiffer in structured meaning, in just the way we have found to matter to the logictutor. So it is plausible that reading (19b) is available for (19) because of thisdifference between it and (19c). (23), which predicts this, is on the right track.

The main point of this article is that (7) should be generalized tosomething like (23), and having made that point, couldn't I stop here? I could ifit weren't for the following remaining loose ends and problems. First, I haven'tbeen able to be very precise yet about the treatment of the other types ofexamples introduced in this section, i.e., (21) and (22). Second, (23) contains anexplicitly disjunctive formulation which it would be nicer to avoid. Third, evenwith the generalized principle (23) in place of the former (7), the theory stillpartitions the phenomena in a strange way: In coreference cases like (14), (23)serves to license a certain interpretation for a grammatical derivation, but incobinding cases like (19) it acts to redeem an ungrammatical one. The actualstatus of the examples does not warrant such a distinction; it would be better ifthey were all grammatical, or all ungrammatical. Or, even better yet, if theywere all somewhere in between, which is what I will actually say below. I willreturn to these points in section 5. The next section is devoted to an independentcriticism of Reinhart's system. The reason why I insert it at this juncture is thatit leads to some technical refinements that will be useful below.

Irene Heim

226

4. From coindexing to linking and codetermination

4.1. Bound variable pronouns that undergo QRA technical difficulty arises when we ask what exactly happens when a boundvariable pronoun undergoes QR. Now I must first convince you that thissituation ever arises in the first place. To be sure, we have made QR completelyoptional and unrestricted, so it would require a special stipulation to prevent itfrom applying to bound variable pronouns. But do we really ever need toexercise this option? The answer is 'yes': If we accept Reinhart's analysis ofellipsis, there will be readings of English sentences that we can only generate byQRing a bound variable pronoun.

This is not the place to launch into a detailed discussion of ellipsis. Iwill just give a very brief exposition of Reinhart's approach and, for simplicity,will concentrate entirely on Bare Argument ellipsis, setting aside any of theadditional complications that arise with the more common and colloquial VPellipsis. Consider a simple ellipsis structure:

(24) I called John, and the teacher too.

(24) is ambiguous: the second conjunct can mean that the teacher called John,or that I called the teacher. In the first reading, the "correspondent" of the"remnant" the teacher is the subject I, in the second reading, the object John isthe correspondent.22 The basic idea, going back at least to Sag 1980, is that anLF for the elliptical conjunct is derived by (a) QRing the correspondent in theantecedent sentence, and (b) inserting a copy of the resulting λ−abstract next tothe remnant. Depending on the intended reading, this procedure yields (24a) or(24b) for (24). (The copied material is in italics.)

(24) a. I λ1[t1 called John], and the teacher λ1[t1 called John] too

b. John λ1[I called t1], and the teacher λ1[I called t1] too

The semantic interpretation of these LFs is transparent.In the two readings of example (24), the correspondents were a

referential pronoun and a proper name. But it is easy to construct similarexamples where the correspondent is a bound variable pronoun. A case in pointis the reading of (25) where it means that every boy said that I called both him(the boy) and the teacher.

(25) Every boy said that I called him, and the teacher too.

To derive the appropriate LF, we must QR the correspondent, which in this caseis the bound variable pronoun him. How exactly does this derivation proceed? 22The terminology of "correspondents" and "remnants" comes from the discussion ofGapping in Pesetsky 1982.


227

First, what's the SS we start out from? Presumably, it has him coindexed withevery boy.

(25) a. SS: every boy1 said that [I called him1, and the teacher too]

(The bracket just serves to indicate that the conjunction is in the lower clause.)On the way to LF, both every boy1 and him1 must QR, the former to bind thepronoun, and the latter to create a constituent for ellipsis copying. So far, wehave assumed that QR applies in such a way as to shift the index on the movedphrase over to the newly created λ next to it, so that the moved phrase itselfwinds up unindexed. And this made sense in view of the fact that indices onquantifying phrases like every boy (as well as indices on proper names and ondeictic pronouns) could make no semantic contribution anyway. However, if weblindly applied QR to him1 in exactly this same fashion, we get somethingundesirable, namely (25b).

(25) b. every boy λ1[t1 said that

[him λ1[I called t1], and the teacher λ1[I called t1] too]]

The him has lost its index to the λ next to it and therefore is no longer bound by(the λ next to) every boy. So (25b) cannot represent the intended reading.

It is not hard to think of ways to avoid this problem. The most obviousand elegant option that comes to mind is to make the disappearance of the indexunder QR simply optional: we are free to either retain the index of the QRedphrase on both itself and the λ, or else—as before—to retain it only on the λ. Inmost cases, we will effectively be forced to the second choice, because wewould otherwise end up with indexed NPs that don't qualify as variables under(5). But nothing prevents us from retaining two copies of the index in thespecial case where we are QRing a bound variable pronoun, and thus we canderive an appropriate LF from the SS (25a), namely (25c).

(25) c. every boy λ1[t1 said that

[him1 λ1[I called t1], and the teacher λ1[I called t1] too]]

This expresses the intended meaning, and so we seem to have solved ourproblem.

But wait, there is a complication: Consider the following slightly morecomplex example:

(26) Every boy said that [he called his mother and the teacher too].

This has many different readings, most of which are to be ignored here. We areonly interested in readings where (a) the correspondent of the the teacher is theembedded subject he, and (b) both he and his are anaphorically related to everyboy. Still, there are two distinct readings that fit these specifications. Thedifference comes out in the following two paraphrases for the elliptical conjunct:

Irene Heim

228

(26) a. ... and the teacher called his own (the teacher's) mother.

b. ... and the teacher called his (the boy's) mother.

(26a) is a sloppy reading and (26b) a strict one. What SS and LF representationsare associated with each of these two readings?

Under our present assumptions, there is only one SS and LF availableto represent either reading: Because both he and his are to be bound variablepronouns and anaphorically related to every boy, we can't but assign thefollowing indexing at SS:

(26) c. every boy1 said that he1 called his1 mother and the teacher too

From there on, we have no real choices (not counting derivations that terminatein uninterpretable LFs or leave the he unbound). Two applications of QR andellipsis copying yield (26d).

(26) d. every boy λ1[t1 said that

[he1 λ1[ t1 called his1 mother]

and the teacher λ1[ t1 called his1 mother] too]]

(26d) represents the sloppy meaning (26a). But for the other reading, the strictone in (26b), we are left without any possible derivation.

4.2. Inner and outer indicesTo remedy this limitation, I propose that we allow pronouns to be doublyindexed at SS already. They can have an inner index that encodes what they arebound by, and an additional outer index to encode what they in turn bind. (Thiswill be made more precise right below.) The inner and outer index need not bethe same.

Such dual indexing may look like a new-fangled notational contrivance,but the distinctions it is meant to express are anything but new. Twoparticularly important precedents are found in the PTQ fragment (Montague1974) and in the Linking framework of Higginbotham 1983. Both of theseprovide two different analyses for a sentence like (27) (a simplified version of(26), without the elliptical conjunct).

(27) Every boy said that he called his mother.

In PTQ, one option (call it (a)) is to build up a sentence with three occurrencesof the same free pronoun: he1 said that he1 called his1 mother, then use S14with the operation F10,1 to quantify in every boy. In this derivation, both surfacepronouns are bound (=lose their subscripts) simultaneously in the last step.Another option (b) would be to generate he2 called his2 mother, then quantifyhe1 into this to yield he1 called his mother, then build up further to he1 said thathe1 called his mother, and finally quantify in every boy with F10,1. This time,


229

the his was bound earlier in the derivation, whereas the he had a subscript untilthe last step. 23 Of course, the meanings are equivalent, as one translationreduces to the other by λ-conversion.

In Higginbotham's framework, (27) is generated with two differentlinked structures:

(27) a. every boy said that he called his mother ↑↑__________| |

|____________________|

b. every boy said that he called his mother ↑_________|↑________|

These correspond respectively to the (a) and (b) derivations in PTQ. Given thesemantics for linked structures in Higginbotham 1987:125–127, 130–131, theyare again logically equivalent.

My double-indexing scheme mimics the two PTQ-derivations (a) and(b) as follows:

(27) c. SS: [every boy]1 said that [he1] called [his1] mother

LF: every boy λ1 [t1 said that he1 called his1 mother]

d. SS: [every boy]1 said that [he1]2 called [his2] mother

LF: every boy λ1 [t1 said that he1 λ2 [t2 called his2 mother]]

The LFs here and their intended semantics should be self-explanatory, but thenew SSs and SS-to-LF-changes call for some comments. Notice the distinctionbetween [αi] (inner index only), [α]i (outer index only), and [αj]i (doublyindexed). Regarding the operation of QR, I return to our original view that italways works the same way. The trace retains the outer index of the movedphrase, but the moved phrase itself transfers it to the λ and thereby loses it.Bound variable pronouns are no exception. If an NP doesn't have an outer indexat SS, it just can't QR. On the other hand, if it does have one, it must QR, or else 23Of course, these are just two out of infinitely many derivations. For one thing, thereare infinitely many isomorphic derivations to each (a) and (b) where 1 and 2 are replacedby different numbers. More interestingly, there are infinitely many additional derivationsnot isomorphic to (a) or (b) which also yield the same meaning. Not all of these can alsobe distinguished by means of my double indices. For instance, in PTQ we might buildhe1 said that he1 called his1 mother, then quantify in he2 with F10,1, and then quantify inevery boy with F10,2. This derivation is not replicable in my system -- unless I were toallow triple indices at SS, but I am not aware of any empirical motivation for this. Thereare, however, some more types of PTQ-derivations besides (a) and (b) that do correspondto distinct double-indices-representations; one of them will in fact become relevant rightbelow, see (26e) and (26g). (This footnote was prompted by questions raised by BarbaraPartee (p.c.). I am aware, of course, that I have not even begun to seriously explore theexpressive capacities of the proposed notation and its relation to alternatives in theliterature.)

Irene Heim

230

its outer index will fail to qualify on the following revised definition of"variable" and the LF will be flitered out. (Note the change in clause (iii).)

(28) Definition of "variable", revised:

An index is a variable only if it is(i) on a λ, or(ii) on a trace and bound by a λ, or(iii) the inner index of a pronominal or anaphor and A-bound.

It follows that a bound variable pronoun that is to QR needs two indices: anouter one in order to QR and an inner one in order to be bound.

Let's return to our problem with the strict-sloppy ambiguity in example(26). Instead of the previous single option in (26c), we now have two choicesfor representing (26) at SS, even when both pronouns are to be bound and thesubject is the correspondent:

(26) e. [every boy]1 said that [he1]2 called [his1] mother and the teachertoo

f. [every boy]1 said that [he1]2 called [his2] mother and the teachertoo

(26f) is, of course, just like (27d) above; (26e) is like (27c) as regards theindexing of his, but has an additional outer index on he (which we need toenable it to QR, a prerequisite for ellipsis copying).24 Each of these derives aunique LF:

(26) g. every boy λ1[t1 said that



h. every boy λ1[t1 said that



And these correctly express the strict and sloppy readings respectively.

24Ignoring the elliptical conjunct, a PTQ derivation that corresponds to the indexingchoices in (26e) would proceed by building up he2 called his1 mother , quantifying in he1by F10,2 to yield he1 called his1 mother, building up to he1 said that he1 called his1mother, and finally quantifying in every boy by F10,1. After the second step, this is likederivation (a), and as in (a), both surface pronouns are bound simultaneously in the laststep.


231

4.3. The strong version of Condition BWe have now solved our initial technical problem about bound variablepronouns undergoing QR, but we have yet to explore the repercussions of oursolution, especially for the operation of the Binding Conditions. With NPsallowed to bear two distinct indices at once, there no longer is a unique obviousnotion of "coindexing," and thus of the Binding Conditions which rely on it.25

A variety of different coindexing-concepts are in principle definable, includingthe following two:

(29) Definition:

a. β is linked to α iff α's outer index = β's inner index.b. α and β are colinked iff α's inner index = β's inner index.

The terminology is deliberately reminiscient of Higginbotham's linkedstructures. It'll make sense to you if you glance at (27a,b) while applying thedefinitions to (27c,d). In fact, linking as defined in (29a) has exactly theproperties of Higginbotham's linking relation; for instance, unlike coindexing itis neither symmetric nor transitive.

Consider now Condition A. In the version used so far, it requires thatan anaphor be "coindexed with" a c-commanding A-position in its GC. Howshould we reinterpret it in the present setting? Higginbotham proposes toreplace "coindexed with" by "linked to":

(30) Condition A, new version:

An anaphor is linked to a c-commanding A-position in its GC.

(30) predicts that an SS such as (31a) is ill-formed.

(31) a. * [he1]2 cut [himself1]

The reflexive, though in some sense "coindexed" with the subject, is not linkedto it in the sense of definition (29a). What (30) requires instead is the kind ofindexing shown in (31b).

(31b) [he1]2 cut [himself2]

The semantic import of this prediction is that a reflexive must really be boundwithin its GC, not just have a cobound antecedent there. To see this, look at theLFs that (31a,b) give rise to:

25The reference to "A-binding", and hence to "coindexing", in LF-conditions like (28)remains unproblematic, since all indices are single by the time we reach a well-formedand interpretable LF.

Irene Heim

232

(31) c. from *(31a): he1 λ2[t2 cut himself1]

d. from (31b): he1 λ2[t2 cut himself2]

The reflexive is free in (31c) and bound in (31d). Neither of these, of course, isa well-formed LF on its own on our current assumptions, because of the freeindex on he1 . But they might be part of larger well-formed LFs such as (32a,b).

(32) a. every boy λ1[t1 said that he1 λ2[t2 cut himself1]]

b. every boy λ1[t1 said that he1 λ2[t2 cut himself2]]

These two happen to be logically equivalent, so we cannot observe right herewhether the prediction that (32a) comes from an ungrammatical SS is borne out.But if we add an elliptical conjunct, the difference becomes manifest: (30)predicts, correctly, I assume26, that (33)—unlike (26)—allows only a sloppyreading.

(33) [every boy]1 said that [he1]2 cut [himself*1/2] and the teacher too.

Now let's turn to Condition B. This used to require that a pronominal not be"coindexed with" any c-commanding A-position in its GC. Should we againreplace "coindexed with" by "linked to"? If we have learned the lessons ofPartee & Bach (1984)27 or Higginbotham (1983; 1985), we know better thanthat. Consider the following potential derivation for sentence (34):

(34) every boy said that he called him

a. SS: [every boy]1 said that [he1]2 called [him1] 26More precisely: if (33) allows a strict reading as well, it does so no more easily than(i).(i) John cut himself, and the teacher too.Whatever relaxation of Reinhart's assumptions (here left untouched) about reflexives andellipsis will accommodate a strict reading in (i) should do so for (33). I don't mean todismiss the issue of strict readings with reflexives or diminish its potential significancefor the goals of this paper, but one thing that seems fairly clear is that we don't want atheory that draws the dividing line between (33) and (i).27I am indebted to Barbara Partee for reminding me of this important reference. It maybe the earliest attempt to formulate Binding Conditions A and B within a frameworkwhich (due to its kinship with PTQ, in this case) captures semantic distinctions that turnon the difference between "cobound" and "bound by", such as the strict/sloppy ambiguityin (26). Just as in Higginbotham's framework, the need to strengthen Condition B arisesas an immediate consequence of the system's ability to represent this semantic distinction.Partee & Bach present the issue somewhat differently than I have done here: Theirtreatment of ellipsis (which focusses on VP Ellipsis rather than Bare Argument Ellipsis)does not rely on quantifying-in of pronouns. But the Derived VP Rule that they employinstead has a similar effect of multiplying the possible derivations for sentences with twopronouns (like (34)). A detailed comparison between their proposals and those ofHigginbotham or the present article must await another occasion.


233

b, LF: every boy λ1[t1 said that he1 λ2[t2 called him1]]

(34b) is logically equivalent to every boy λ1[t1 said that he1 called him1], areading that sentence (34) clearly lacks and that English can only express withthe reflexive: 'every boy said that he called himself '. So this derivation shouldbe ruled out. If Condition B merely prohibits linking , it will not block (34a):[him1] is not, after all, linked to [he1]2 (nor is there anything else we know ofwrong with the derivation). [he1]2 and [him1] are, however, colinked (asdefined in (29b)). So if we make Condition B sensitive to both linking andcolinking, we would seem to solve our problem. This suggests the followingformulation:

(35) Condition B, preliminary new version:

A pronominal is neither linked to nor colinked with any c-commandingA-position in its GC.

A noteworthy implication of such a conception of Condition B is that manyindexing patterns cannot be realized with any lexical item at all. For instance,[he1] called [α1] is ill-formed however we spell out α, in particular for α = himas well as α = himself.

(35) is adequate to deal with (34), but it is known to be still too weakfor similar examples with somewhat more complex indexing patterns, forinstance (36).28

(36) * [every boy]1 said [he1]2 knew [he1] called [him2]

The LF that (36) gives rise to has the meaning of 'every boy said he knew hecalled himself ', so this better not be a well-formed SS. Treading again inHigginbotham's footsteps29, I therefore define an even more inclusive relationof which linking and colinking are two, but not the only, special cases.

(37) Definition30:

28This too was already seen clearly by Partee and Bach (1984). See especially theirsection 6.3, pp. 124 - 25, where they comment on the sentence Every man talked tohimself by himself about a book that he gave to him and the need for a condition that issensitive to global properties of the derivation.29See especially the motivation for the "transitivity condition" in Higginbotham 1983:406 and the refined conditions that replace it in 1985: 570 - 575 and 1991: 8 - 11. Thefollowing is an attempt to replicate his proposal that two elements in a Condition Bconfiguration "cannot be determined by the structure in which they occur to share avalue" (1985: 575).30This is an equivalence relation. In the present system, it happens to be coextensivewith the reflexive, symmetrical, and transitive closure of the linked-to-relation. It willdiffer, however, after we reintroduce free variables in section 5.2 below. Two NPs with

Irene Heim

234

α and β are codetermined iff(i) α = β, or(ii) either one of α or β is linked to the other, or(iii) α and β are colinked, or(iv) for some γ, α and γ are codetermined and so are γ and β.

(38) Condition B, final new version:

A pronominal is not codetermined with any c-commanding A-position inits GC.

I leave it to the reader to check this against (36) and other even fancier casessuggested by the literature. For the remainder of this article, it is sufficient toremember that codetermination includes linking and colinking.

To sum up section 4, I started out with the question of how QR appliesto bound variable pronouns and ended up with an endorsement ofHigginbotham's Linking Theory, in particular of its non-parallel conceptions ofBinding Conditions A and B. The overall indexing system appears morecomplex now, but it may be seen as a conservative extension of Reinhart's ideas.Notice in particular that, throughout my revisions, I have adhered to herassumption that there is a perfect correspondence between syntactic binding inthe sense of Condition A and semantic binding (variable binding at LF). Thenewly introduced double indices serve, in effect, to express distinctions at SSthat otherwise could only be brought out at LF. And the motivation forexpressing them there is (as throughout Reinhart's work) the fact that syntacticBinding Conditions, which apply at SS, are sensitive to those distinctions.

5. Revised proposal

5.1. The Exceptional Coindexing Rule

We are now ready to return to the central issue of this study, namely the statusof the Coreference Rule. In the section 3 titled "reference isn't special", Isuggested that whatever will survive of this rule should pertain not just tocoreference possibilities but also to the possibility of cobinding two boundvariable NPs. Consider again the derivation I gave above for the donkeysentence version of the logic tutor's utterance and which violated (the originalversion of) Condition B:

(19) a. SS: * if everyone hates [a man]1,

[that man himself]1 hates him1

b. LF: ∀1 [ if [a man]1 [everyone hates t1] ] matching free inner indices will then count as colinked by (29b), and hence ascodetermined by (37iii), and this is intended.


235

[ [that man himself]1 hates him1 ]

From our current perspective, the question arises whether the index on that manhimself is supposed to be an inner or an outer index. Given that it encodes whatthis bound variable demonstrative is bound by , it makes sense to consider it aninner index, as in (19d,e).31

(19) d. SS: * if everyone hates [a man1],

[ that man himself1] hates [him1]

e. LF: ∀1 [ if a man1 [everyone hates t1] ]

[ that man himself1 hates him1 ]

This derivation, at any rate, seems appropriate to the intended reading. Noticehow (19e) illustrates Universal Instantiation: under any given assignment, thepredicates following everyone in the antecedent and that man himself in theconsequent denote the same property.

Our worry above was that (19a) was excluded by Condition B. (19d) isstill excluded by the new version of Condition B: [that man himself1] and[him1] are colinked. If we say nothing further, we are no better off than before.Suppose, however, we exploit the refined indexing system to draw a distinctionbetween different types of Condition B violations: linking violations and merecodetermination violations. In the former, a pronominal is linked to a c-commanding A-position in its GC; in the latter, it is codetermined with one, butnot linked to any. Notice that [him1] in (19d) is codetermined with [that manhimself1], but not linked to it, so this is a mere codetermination violation. Ispeculate that this is significant. The general hypothesis is as follows: Linkingviolations of Condition B are disallowed under any circumstances. Merecodetermination violations are normally disallowed as well, and perhaps arenever 100% perfect, but they can be much improved under suitable discourseconditions. Specifically, what makes (19) viable could be a principle like thefollowing.

(39) Exceptional Coindexing Rule:

A pronominal α is (marginally) allowed (at SS) to be codetermined with ac-commanding A-position β in its GC when the interpretation thusobtained needs to be distinguished from the one that would result if (β

31The pertinent clauses in the definition of variables (cf. footnote 21) are accordinglyreinterpreted as follows:(iv) or the inner index of an indefinite in the restriction of a QAdv and A-bar-bound

by it,(v) or the inner index of a definite in the nuclear scope of a QAdv and A-bar-bound

by it.

Irene Heim

236

were given an outer index and moved and) α were replaced by a variableA-bound by the trace of β.

This is obviously a very close relative of Reinhart's Coreference Rule. Not onlydoes it express her basic insight that we may sometimes say things we otherwisecouldn't because that's the best way to get our communicative point across; italso retains almost all the concrete wording of (7). But one thing is different:(39) is about indices, not about referents.

(39) applies to sanction (19d,e), on the grounds that the meaning of thecompeting structure in (19f) would not have served the logic tutor's specialcommunicative purpose:

(19) f. ∀1 [ if a man1 [everyone hates t1] ]

[ that man himself1 λ2[t2 hates him2] ]

Though logically equivalent, the structured meaning of (19f) fails to illustratepure Universal Instantiation.

Now it would be rather dissatisfying to have both the Coreference Rule(7) and the Exceptional Coindexing Rule (39) side by side in our system. Theyclearly concern different instances of the same phenomenon and we would liketo collapse them somehow. But how is this possible? Under the presentassumptions about syntactic representation which we took from Reinhart, thereis a fundamental difference between (7) and (39): (39) deals with licensingcertain indexing patterns, but (7) doesn't deal with any aspect of syntacticrepresentation at all, but directly with interpretive choices. The reason for thisis, of course, Reinhart's view that coreference relations, as opposed to variablebinding relations, have no syntactic representation. If we want to unify (7) and(39), it seems advisable to give up this radical bifurcation. In particular, if wewant to subsume the cases now dealt with by (7) under the scope of (39), we aredriven back to the more conventional view of coindexing as a representationaldevice that does double duty for coreference and variable binding. This is theline that I will pursue here, and it requires a couple of revisions in the basicsyntactic and semantic system.

5.2. Free variables readmittedIf referring NPs are to bear indices and indices are semantically variables, thenreferring NPs correspond to free variables. Free variables were systematicallyexcluded in Reinhart's system, so our first task is to permit free variables underthe well-formedness conditions for LFs. The current definition of "variable"leaves no room for them, so we have to extend it.

(40) Definition of "variable", revised again:

An index is a variable only if it is(i) on a λ, or(ii) on a trace and bound by a λ, or


237

(iii) the inner index of a pronominal or anaphor and A-bound, or(iv) the inner index of a definite NP and free.

Second, we must provide these free variables with some kind of semanticinterpretation. Let's at first restrict our attention to the special case where we aredealing with free pronouns . These are, of course, the deictic pronouns (formerlyunindexed) and they get their reference assigned by the context. Since we wantto ensure that coindexed free pronouns are ipso facto coreferential, we shouldnow think of the context as supplying a referent per index (rather than, as before,per pronoun occurrence). And mindful of our discussion of cognitive value insection 2.3, we assume that it is not really a referent that the context supplies,but a guise. So we add the following principle:

(41) Reference Assignment (1st version):

An LF φ is interpretable in an utterance context c only iffor each index i free in φ, c furnishes a guise Fci.

In the compositional semantic interpretation of the utterance, these contextuallysupplied guises provide values for the free variables. When we are calculatingthe proposition expressed by an LF φ in a given context c, we assign to each freeαi the individual Fci(wc) that Fci picks out in the utterance world wc. (To obtainthe cognitive value, we first have to calculate the propositional concept thatmaps each world w in the context set of c to the proposition φ would express ifuttered in w, then diagonalize this; see Stalnaker 1979.)

As it stands, (41) ensures that coindexed free pronouns corefer. In fact,it ensures something even stronger, namely that they pick out the same referentunder the same guise . How about two free pronouns with different indices?Could they still receive the same referent? Could they be assigned even thesame guise? Nothing so far prevents it, and as far as mere coreference goes,that's just fine. But sameness of guises is something we have to disallow, forreasons that will quickly become clear. So we need an amendment:

(42) Reference Assignment (2nd version):

An LF φ is interpretable in an utterance context c only if c furnishes a distinct guise Fci for each index i free in φ.

With this much of the new system in place, the coverage of our ExceptionalCoindexing Rule is no longer limited to bound variable examples like (19), butincludes analogous cases with two coreferential pronouns, such as the followingvariant of (14).

(43) Look, if everyone hates Oscar, then it surely follows that he himself hates him.

Suppose the intended interpretation of this has the following derivation:

Irene Heim

238

(43) a. SS: if everyone hates Oscar, ... [he himself2] hates [him2]

b. LF: if everyone hates Oscar, ... he himself2 hates him2

The utterance context c for (43b) is of course meant to be such that Fc2(wc) =Oscar. (Never mind what the complete appropriate guise of Oscar is.) So theconsequent clause of (43b) in context c expresses the structured proposition thatconsists of Oscar and the property of hating Oscar. Now (43a) is a Condition Bviolation because [he himself2] is in the GC of and codeterrmined with [him2].But (39) allows it anyway because this is a context where structured meaningmatters, and the alternative (43c) would express a different structuredproposition in c, namely the one consisting of Oscar and the property of self-hating.

(43) c. ... he himself2 λ3[t3 hates him3]

So far, so good. Our account of this example is not complete, however, withouta demonstration that (43a,b) is essentially the only derivation under which it getsthe intended meaning (apart from trivial alphabetic variants, of course). Inparticular, we have to watch out for the following two potential rivals, in whichhe himself has either received no index at all, or a different one from him.

(43) d. SS: if everyone hates Oscar, ... [he himself] hates [him2]

e. LF: if everyone hates Oscar, ... he himself hates him2

(43) f. SS: if everyone hates Oscar, ... [he himself4] hates [him2]

g. LF: if everyone hates Oscar, ... he himself4 hates him2

Neither of these are even prima face Condition B violations, so anyinterpretations they represent should be fine quite regardless of the specialdiscourse conditions that are required for clearance by the ExceptionalCoindexing Rule. They better not be alternate representations of our intendedinterpretation, then, because that way we would lose our empirical predictionabout (43). In other words, we would lose Reinhart's insight that this particularinterpretation of this English sentence stands and falls with the presence of acontext in which there is some special reason for structured meaning to matter.

As regards (43d,e), I will simply stipulate it away by a provision thatpronouns bear an (inner) index as a lexical property (see right below). (43f,g),on the other hand, must certainly be allowed as a well-formed derivation forsome reading of (43), for example one where he himself refers to Fred. But canit be set in a context c where it so happens that Fc4(wc) = Fc2(wc) = Oscar?Well yes, it could so happen —for instance if we have been talking about acertain "Fred", he himself4 connects anaphorically to an occurrence of that nameFred, and then Fred and Oscar turn out to be the same person. But in this case,Fc4 ≠ Fc2 (even though they concide on wc), and that's a different interpretationfrom the one we have been contemplating all along (and one that should bepredicted okay without special excuses about the importance of structured


239

meaning!). What our Reference Assignment principle doesn't allow is for hehimself4 and him2 in (43g) to pick out Oscar through the same guise. (This wasthe purpose of the amendment in the 2nd version.) And this, I take it, suffices todisqualify (43f,g) as an alternative representation of the interpretation weassociated with (43a,b). The logic tutorial setting that we have been imaginingjust doesn't provide two separate guises of Oscar (much less appropriate clues todetermine which pronoun goes with which guise).

So much for (43). But of course we also need to cover the original(14), with the proper name Oscar (himself) rather than a pronoun in the subjectposition. In order to do this, we must allow the name to bear an (inner) indexand that index must represent the name's reference. What's more, we must notjust allow, but force , names to bear indices. This is for the same reason that wewanted to exclude (43d,e) above: If we let a name be interpretable without thebenefit of an index, we couldn't read off coreference relations between that nameand some other NP from the syntactic representation. That way, we couldalways "sneak by" Condition B and the Exceptional Coindexing Rule bysuppressing potentially offending indices and still get equivalent readings. Andthen we would still need a separate Coreference Rule of Reinhart's kind to blockthose.

So we must go all the way and make sure that there is no referencewithout an index . Specifically, I stipulate that all pronouns, proper names, andother definites bear an inner index as part of their lexical entry and thus comeautomatically indexed when they are inserted in any syntactic structure. (Theassignment of outer indices may be left free and optional as before.)

Free indices still are assigned referents by the utterance context, but theNP's lexical content must now be made to constrain this assignment. Actually,to some extent this happens even with pronouns, which normally have gender,number, and person features that limit their choice of referent. Proper names arejust a more extreme case of this phenomenon: he fits any male person or animal,John is specialized to those named "John". Since that's a common name, there isstill considerable disambiguating work left to be done by the usual pragmaticfactors; but the formulation below is also meant to cover the limiting case of aname unique enough to render the context's role trivial.

(44) Reference Assignment (3rd version):

An LF φ is interpretable in an utterance context c only if c furnishes adistinct guise Fci for each index i free in φ

such that it is presupposed in c that Fci picks out an individualwhich fits the features and lexical content of any NP indexed i occurringin φ.

(44) is a bit vague, but that could be worked out.32

There is now no significant difference between (43) and the original(14). On our current assumptions, the latter's SS and LF are as follows:

32For some pertinent discussion, especially regarding definite descriptions, see Heim1988.

Irene Heim

240

(14) c. SS: if everyone hates [Oscar2], ... [Oscar2] hates [him2]

d. LF: if everyone hates Oscar2, ... Oscar2 hates him2

The context must match the free index 2 with some guise that is known to be ofsomething named "Oscar" and a male human or animal. (If the secondrequirement, stemming from the features of the pronoun, is redundant, given that"Oscar" is a familiar name for males, so much the better.) Presumably, thisguise is the memory entry under the name "Oscar".33 The ExceptionalCoindexing Rule applies to license this interpretation just as we showed abovefor (43a,b), and (for the reasons we went over there) we needn't worry aboutalternate derivations in which Oscar and him weren't coindexed.

5.3. Other examples in the revised analysisIn arguing that the Exceptional Coindexing Rule takes over the work ofReinhart's Coreference Rule, I have looked at only one type of example so far,the kind where structured meaning matters. What about the other types initiallyadduced in illustration of the Coreference Rule? Let's run through the list onemore time. We will see that the revised account parallels G & R's original onefor the only and Lakoff examples. On the other hand, due to the fact that (39)doesn't apply unless there is a Condition B violation in the first place, the basiccases and debates about identity are now licensed independently of (39), simplybecause they have fully grammatical representations.

5.3.1. Basic casesHere is our initial trio again:

(8) John saw him .


(10) His mother saw John .

We can be brief on (8): A bound variable interpretation is out as before, as itwould be a linking violation of Condition B. A corefential interpretation, withthe common referent picked out under the same guise, would have to berepresented by matching inner indices on John and him and thus would be amere codetermination violation of B. It is therefore also out, except in specialcontexts where structured meaning matters. Finally, coreference via distinctguises is okay, but requires the sort of context where distinct guises are readilyavailable.

With (9), Condition B does not apply since John is beyond the GC ofhis. Both linking and colinking are therefore permitted. The former representsthe bound variable reading, the latter the coreferential one. Here we see thebenefit of restricting (39) to prima facie ill-formed structures (as opposed to 33In the logic tutorial setting, there may be no reference to any particular real-life Oscar.But let's abstract away from this bit of make-believe.


241

haveing it apply everywhere, as did (7) and (23)). As G & R point out in theirfootnote 13, there is reason to believe that (10) is inherently ambiguous.Otherwise one would have to say that it allows a coreferential interpretation incontinuations like John saw his mother, and so did Bill only because this yieldsa manifestly distinct interpretation for the continuation. But this is problematicfor two reasons. First, as noted in Lasnik 1989:166, it doesn't straightforwardlycover the case where the elliptical continuation is uttered by a different speaker.Second, there is independent reason to doubt that the Coreference Rule "looksahead" that far in the discourse: if it did, John saw him, and Bill did too, shouldlikewise permit coreference between John and him. G & R show that theseproblems disappear if the Coreference Rule is restricted to prima facie ConditionB violations, and I have taken up a version of that idea in my (39).34

In (10), variable binding continues to be out for the same reason asbefore: Locally A-bar-bound pronouns still fall through the cracks of thedefinition of "variable". But nothing whatsoever interferes with coreference,which would be represented by matching inner indices on his and John. (Again,I mean coreference via a common guise.)

5.3.2. Identity debatesLet's turn to example (12).

(12) A: Is this speaker Zelda?

B: How can you doubt it? She praises her to the sky. No competingcandidate would do that.

In a way, I have already said everything I have to say about this one. Withoutbeing mentioned explicitly at that point, this type of example was my guide insetting up the Reference Assignment principle in such a way that it imposes aone-to-one match of indices and guises (and not of indices and actual referents ).So whether or not party B turns out to be right and the speaker is Zelda, thesyntactic representation of the pertinent reading of this example will havedistinct inner indices on she and her, and so, once again, Condition B doesn'teven get a shot at it.

Here the present analysis actually takes a different view than G & R,even in the version of their footnote 13, which restricts (7) to prima facieCondition B violations. (12) does qualify as such a prima facie violation in theirsense, but it isn't a violation of any kind in the present system. This is welcomeinsofar as this type of case has none of the marginal flavor of the other types (asG & R also concede in their footnote 10).

5.3.3. only examplesHere is the example about Lucifer once more:

(11) (Everybody hates Lucifer.) Only he himself pities him . 34G & R actually don't commit themselves to this particular solution, they just spell it outas one possibility among others.

Irene Heim

242

Recall that there was a murky point in G & R's original account of this example,a kind of equivocation between "the trace of only he himself" and "the trace ofhe himself". Let us see what shape this takes in the present system. Thederivation of (11) on the intended reading ought to be (11c,d).

(11) c. SS: [only [he himself1]]2 pities [him1]

d. LF: only [he himself1] λ2 [t2 pities him1]

As indicated, 1 in (11c) is an inner index on the smaller NP, and 2 an outer index on the larger NP. If this is indeed the situation, then (11c) is not inviolation of Condition B, not even a codetermination violation: the larger NP isin a c-commanding A-position w.r.t. [him1] but doesn't share any index with it,and the smaller one is colinked with [him1] but not in a c-commanding A-position.

If we are only concerned with allowing this interpretation of (11), wecan live with this result. But the intuitive status of this type of example is likethat of others we have classified as weak Condition B violations in need ofsanctioning by the Exceptional Coindexing Rule. So a better prediction wouldbe that the subject and object in the SS of (11) do count as codetermined, andthat therefore (39) has to step in to license (11d) on the grounds that its truth-conditions differ from those of (11e):

(11) e. only [he himself1] λ2 [t2 pities him2]

At this point, our options are not unlike G & R's: For instance, we can juststipulate that adjunction structures of this sort count as one NP rather than two.(This implies, among other things, that such structures never have room formore than a total of two indices; they couldn't have up to four, as we'd expect ina genuine instance of two nested NPs.) The indices 1 and 2 on he himself andonly he himself then just are the inner and outer index of this one NP, whichthereby qualifies as colinked with the object. As it stands, this is just as hokeyas the equivocation implicit in G & R's original story. But perhaps it can beimproved upon when future research gives us a deeper understanding of both thesyntax and the semantics of double indexing.35

35In a related type of example that is likely to be relevant in this connection, only doesnot form a constituent with the focus it associates with:(i) I could only see that [he himself]F pities him.As shown by Rooth (1985: 72 - 72, 76 - 79), an index on a focussed phrase may be insideor outside the focus-marker, with non-equivalent semantic results. In the present setting,we face the question whether the phrase including the F and the one that excludes itshould be treated as two separate, nested NPs (each with potentially an inner and an outerindex), or whether there is just one NP, whose inner index is interpreted below the F andwhose outer index above it. The latter choice is preferable in light of the fact that therelevant coreferential readings for such sentences have the same intuitive status asanalogous cases in which only and its focus form a constituent. -- The discussion ofanaphora with focussed antecedents in Higginbotham 1989 also points in this direction.


243

In other words, our analysis of (11) remains essentiallyReinhart's—except that it now extends automatically to (21), which receives thederivation below:

(21) d. SS: [every devil]1 knows that

[only [he himself1]]2 pities [him1]

e. LF: every devil λ1[t1 knows that

[ only [he himself1] λ2[t2 pities him1] ] ]

The fine-structure of the only-NP in (21d) is whatever it is in (11c), and thus(11) and (21) are alike from the standpoint of both Condition B and theExceptional Coindexing Rule. The potential preempting alternative to (21e),which however has different truth-conditions, is (21f).

(21) f. every devil λ1[t1 knows that

[ only [he himself1] λ2[t2 pities him2] ] ]

5.3.4. Lakoff's exampleG & R's account of (15) likewise carries over in full and now generalizes to (22).The derivations speak for themselves.

(15) i. SS: [I2] dreamt [I2]1 kissed [me2]

j. LF: I2 dreamt λ1 [ t1 kissed me2 ] [ I2 ]

k. potential preemptor: I2 dreamt λ1 [ t1 kissed me1 ] [ I2 ]

(22) b. SS: [not only [I3]]2 dreamt [I2]1 kissed [me2]

c. LF: not only [I3] λ2 [t2 dreamt λ1 [ t1 kissed me2 ] [ I2 ] ]

d. potential preemptor:

not only [I3] λ2 [t2 dreamt λ1 [ t1 kissed me1 ] [ I2 ] ]

6. Concluding remark

I have pointed out plenty of loose ends along the way. Let me just note onemore respect in which the present account might need revision. My proposal, asit stands, actually has two separate parts. One claim is that there are twodifferent types of Condition B violation, and that one of these types, the mere

Irene Heim

244

codetermination violations, can be rescued under suitable discourse conditions.The second claim specifies more concretely what these discourse conditions are;in close agreement with Reinhart, it implies that an interpretive differencebetween the LF to be licensed and its binding-alternative is a sufficient licensingfactor.

Now it could conceivably turn out that the first claim is right, but thesecond is wrong. Specifically, there may be additional discourse factors, besidesor even instead of the one that Reinhart identified. For instance, Evans (1980)and others have conjectured that prima facie Condition B violations improvewhen the previous discourse provides an antecedent that both potentiallyoffending NPs can be anaphorically linked to. G & R correctly observe that thisis not a sufficient condition for such violations to become acceptable. But apossibility that has not been explicitly considered so far is that it might be anecessary factor in addition to the requirement stated in (7) or (39). There issomething to be said for this possibility. All the relevant examples I have usedin this article actually happen to have this property; see, in particular, (11), (14),(15), and their variants, all of which contained a suitable common antecedent forthe two NPs in the Condition B configuration. As Hans Kamp (p.c.) pointed outto me, this appears to be crucial. For instance, if we have not been talking aboutLucifer before, an utterance of Only Lucifer pities him (or of Only he pities him,using the first pronoun with a demonstrative reference to Lucifer) is quite badunder a coreferential reading. So the condition identified by Reinhart andcarried over into my Exceptional Coindexing Rule is not by itself sufficient, anda common antecedent seems to be required on top of it. This needs furtherexploration.

References

Black, Max and Peter Geach eds. (1952) Translations from the Philosophical Writings ofGottlob Frege, Oxford: Blackwell.

Chierchia, Gennaro (1991) "Anaphors and Attitudes De Se," in R. Bartsch, J. vanBenthem, P. van Emde Boas eds. Language in Context, Dordrecht: Foris.

Cresswell, Maxwell and Arnim Stechow (1982) "De Re Belief Generalized," Linguisticsand Philosophy 5.4, 503 -535.

Dalrymple, Mary, Stuart Schieber, Fernando Pereira (1991), "Ellipsis and Higher-OrderUnification," Linguistics and Philosophy 14.4, 399–452.

Evans, Gareth (1980) "Pronouns," Linguistic Inquiry 11, 337–362.Fiengo, Robert and Robert May (1994) Indices and Identity, MIT Press, Cambridge.Frege, Gottlob (1892) "Über Sinn und Bedeutung," Zeitschrift für Philosophie und

philosophische Kritik 100, 25–50.Grodzinsky, Yosef and Tanya Reinhart (1993) "The Innateness of Binding and

Coreference," Linguistic Inquiry 24, pp. 69–102.Haas-Spohn, Ulrike (1995) Versteckte Indexikalität und subjektive Bedeutung, Berlin,

Akademie Verlag, Studia Grammatica 38.Haïk, Isabelle (1984) "Indirect Binding," Linguistic Inquiry 15.2, 185–223.Heim, Irene (1988) The Semantics of Definite and Indefinite Noun Phrases, New York:

Garland Press. [reprinted from 1982 Ph.D. thesis, Univ. of Massachusetts,Amherst]

Heim, Irene (1990) "E-Type Pronouns and Donkey Anaphora," Linguistics andPhilosophy 13, 137–177.


245

Higginbotham, James (1980) "Anaphora and GB: Some Preliminary Remarks," in J.Jensen ed. Cahiers Linguistiques d'Ottawa: Proceedings of NELS 10, Univ. ofOttawa.

Higginbotham, James (1983) "Logical Form, Binding, and Nominals," Linguistic Inquiry14, 395–420.

Higginbotham, James (1985) "On Semantics," Linguistic Inquiry 16, 547–593.Higginbotham, James (1987) "On the Varieties of Cross-Reference," in GLOW X

Proceedings, Annali di Ca'Foscari, Venice, 123–142.Higginbotham, James (1989) "Reference and Control," Rivista di Linguistica 1.2,

301–326.Higginbotham, James (1994) "Anaphoric Reference and Common Reference," Linguistic

Inquiry.Kaplan, David (1989) "Demonstratives," in J. Almog, J. Perry, H. Wettstein eds. Themes

from Kaplan, Oxford: Oxford Univ. Press. [circulated in manuscript form since1977]

Kitagawa, Yoshihisa (1991) "Copying Identity," Natural Language and LinguisticTheory 9.3, 497–536.

Lakoff, George (1972) "Linguistics and Natural Logic," in D. Davidson and G. Harmaneds. Semantics of Natural Language, Dordrecht: Reidel, 545–665.

Landman, Fred (1986) Towards a Theory of Information, Dordrecht: Foris.Lasnik, Howard (1989) Essays on Anaphora, Dordrecht: Kluwer.Lewis, David (1979) "Attitudes De Dicto and De Se," The Philosophical Review 88,

513–543.Lewis, David (1984) "Individuation by Acquaintance and by Stipulation," in F. Landman

and F. Veltman eds. Varieties of Formal Semantics, Dordrecht: Foris, 219–243.McKay, Thomas (1991) "Representing De Re Beliefs," Linguistics and Philosophy 14.6,

711–739.Montague, Richard (1974) "The Proper Treatment of Quantification in Ordinary

English," in R. Thomason ed. Formal Philosophy, New York: Academic Press,247–270.

Neale, Stephen (1990) Descriptions, Cambridge: M.I.T. Press.Partee, Barbara and Emmon Bach (1984) "Quantification, Pronouns, and VP Anaphora,"

in J. Groenendijk, T. Janssen, M. Stokhof eds. Truth, Interpretation andInformation, Dordrecht: Foris, 99–130. [reprinted from Formal Methods in theStudy of Language, Univ. of Amsterdam, Mathematical Centre Tracts. 1981]

Pesetsky, David (1982) Paths and Categories, Ph.D. thesis, M.I.T., Cambridge.Quine, W. V. O. (1956) "Quantifiers and Propositional Attitudes," The Journal of

Philosophy 53, 177–187.Reinhart, Tanya (1983a) Anaphora and Semantic Interpretation, Chicago: The Univ. of

Chicago Press.Reinhart, Tanya (1983b) "Coreference and Bound Anaphora: A Restatement of the

Anaphora Questions," Linguistics and Philosophy 6, 47–88.Reinhart, Tanya (1991a) "Binding and Coreference Revisited," ms. Tel Aviv Univ.Reinhart, Tanya (1991b) "Self-Representation," ms. Tel Aviv Univ. [presented at

anaphora workshop, Princeton Univ., Oct. 1990]Rooth, Mats (1985) Association with Focus, Ph.D. thesis, Univ. of Massachusetts,

Amherst.Sag, Ivan (1980) Deletion and Logical Form, New York: Garland Press. [reprinted from

1976 Ph.D. thesis, M.I.T., Cambridge]Soames, Scott (1990) "Pronouns and Propositional Attitudes," The Proceedings of the

Aristotelian Society.Stalnaker, Robert (1979) "Assertion," in P. Cole ed. Syntax and Semantics 9: Pragmatics ,

315–332.

Irene Heim

246

Stechow, Arnim von (1982) "Structured Propositions," Univ. of Konstanz, report #59 ofthe Sonderforschungsbereich 99.

Stowell, Timothy (1987) "Adjuncts, Arguments, and Crossover," ms. Univ. of California,Los Angeles.

Irene HeimDepartment of Linguistics and PhilosophyE39-245MITCambridge, MA 02139U.S.A.

Documents

Anaphora and Semantic Interpretation: A Reinterpretation of